Compositional Convolutional Neural Networks: A Deep Architecture with\n Innate Robustness to Partial Occlusion

2020/03/09 by Adam Kortylewski, Kortylewski, Adam, Ju He +6 · 2 citations
Computer Science · Medicine · #Advanced Neural Network Applications #COVID-19 diagnosis using AI #Computer Vision and Pattern Recognition (cs.CV) #Digital Imaging for Blood Diseases #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences

paper · pdf · doi:10.48550/arxiv.2003.04490

openalex publication_date 2020/03/09 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28

Abstract

Recent findings show that deep convolutional neural networks (DCNNs) do not\ngeneralize well under partial occlusion. Inspired by the success of\ncompositional models at classifying partially occluded objects, we propose to\nintegrate compositional models and DCNNs into a unified deep model with innate\nrobustness to partial occlusion. We term this architecture Compositional\nConvolutional Neural Network. In particular, we propose to replace the fully\nconnected classification head of a DCNN with a differentiable compositional\nmodel. The generative nature of the compositional model enables it to localize\noccluders and subsequently focus on the non-occluded parts of the object. We\nconduct classification experiments on artificially occluded images as well as\nreal images of partially occluded objects from the MS-COCO dataset. The results\nshow that DCNNs do not classify occluded objects robustly, even when trained\nwith data that is strongly augmented with partial occlusions. Our proposed\nmodel outperforms standard DCNNs by a large margin at classifying partially\noccluded objects, even when it has not been exposed to occluded objects during\ntraining. Additional experiments demonstrate that CompositionalNets can also\nlocalize the occluders accurately, despite being trained with class labels\nonly. The code used in this work is publicly available.\n

Citations

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• ImageNet classification with deep convolutional neural networks
• Unsupervised learning of object semantic parts from internal states of CNNs by population encoding
• Improved Regularization of Convolutional Neural Networks with Cutout
• Visual Concepts and Compositional Voting
• CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features
• TDAPNet: Prototype Network with Recurrent Top-Down Attention for Robust Object Classification under Partial Occlusion

Cited by

• Improving Classification of Occluded Objects through Scene Context
• TOTNet: Occlusion-Aware Temporal Tracking for Robust Ball Detection in Sports Videos

Related

• Compositional Convolutional Neural Networks: A Robust and Interpretable\n Model for Object Recognition under Occlusion
• Combining Compositional Models and Deep Networks For Robust Object Classification under Occlusion
• Localizing Occluders with Compositional Convolutional Networks
• Robust Object Detection under Occlusion with Context-Aware CompositionalNets
• Teaching Compositionality to CNNs